Latest revision as of 10:44, 18 May 2026

Abstract

Point cloud registration is essential for closed-loop digital forming of ship plates, where CAD-toscan alignment is required for surface error evaluation and compensation. Industrial ship-plate point clouds acquired by structured-light or laser sensing are often corrupted by boundary-related structural outliers and are feature-sparse, causing classical ICP to be sensitive to mismatches and slow to converge. We propose a training-free, deployment-oriented pipeline that combines density-based outlier removal with an Andersonaccelerated ICP (A-ICP) formulated in the SE(3) Lie algebra. Experiments on three representative plate geometries show 37.7%–39.5% lower registration error and indicate improved convergence behavior relative to classical ICP variants. The method is further validated on an SKWB-2500 machine-in-the-loop workflow, achieving MAE of 0.68 mm (sail-shaped) and 0.38 mm (saddle-shaped), with corresponding RMSE of 0.7301 mm and 0.4141 mm. Learning-based baselines are not benchmarked due to the lack of a fair in-domain dataset and retraining protocol under proprietary sensing conditions.

Document